VeFed: a virtual epidemiological environment for federated training of clinical language agents

Introduction Federated learning (FL) lets hospitals train shared clinical models without pooling patient data, yet FL is almost always tested on static, pre-split datasets. In the real world, cases arrive as a stream whose volume and disease mix follow each site's epidemic phase, and new diseases can appear that no site has seen. Methods We built VeFed, a testbed that couples stochastic SIR epidemics or controlled case-arrival schedules with LLM-generated patient-clinician dialogs across hospital silos. Each silo fine-tunes a LoRA-adapted DistilBERT classifier, and we compared six aggregation rules (FedAvg, FedProx, FedSGD, FedAdam, FedNova, SCAFFOLD). Fictional diseases ruled out prior knowledge from pre-training. We replicated key tests on MIMIC-IV-ED and MIMIC-IV-Note text and tested each claim against five controls fixed before they were run. We also introduce PrototypeBank, a federated nearest-centroid method that detects and names new diseases without retraining. Experiments used 10 seeds unless noted. Results Case-arrival shape changed time to competence (round-10 accuracy 0.557–0.920 across five shapes) but not final accuracy on synthetic text. On real clinical notes, the ramp schedule also lowered final accuracy (0.721 vs. 0.885 for burst). Staggered epidemic peaks across silos caused no detectable penalty. With ample data, federated and isolated training reached the same clustering quality and detected a new disease at the same round. With scarce data, prior federation raised novel-disease recall to 0.991, compared with 0.704 for isolated training ( n = 3 seeds). Softmax-based detection flagged known diseases as novel 92% of the time, and PrototypeBank cut this to 16%. kNN detection still recovered more novel cases (recall 0.954 vs. 0.444). Detection latency held from 3 to 10 silos (10.9 vs. 11.2 rounds). LLM paraphrasing cut the mutual information between text and disease label about 15-fold. Discussion Federation value depends on data volume. It adds little when silos have enough cases alone and helps silos where cases are still scarce. PrototypeBank gives up some recall in exchange for a smaller privacy footprint than calibration-set detectors. The proposed simulation provides a reproducible testbed for studying methodological and systemic aspects of FL in healthcare AI, supporting the development of secure, adaptive, distributed clinical intelligence systems.